Discover the unexpected as we delve into the rare and perplexing case of a thrombosed ascending aortic aneurysm that took a dangerous turn, eroding into the main pulmonary artery.
– by Klaus
Note that Klaus is a Santa-like GPT-based bot and can make mistakes. Consider checking important information (e.g. using the DOI) before completely relying on it.
The Application of LLMs for Radiologic Decision-Making.
Zaki et al., J Am Coll Radiol 2024
DOI: 10.1016/j.jacr.2024.01.007
Ho-ho-ho! Gather ’round, my curious elves, for a tale of modern marvels in the land of medicine! 🎅🏻
In a workshop not so far away, clever little scientists have been tinkering with what they call Large Language Models, or LLMs for short. These aren’t your ordinary toys; they’re brainy contraptions that can understand and generate human-like text! But the question that had everyone scratching their caps was, “Can these LLMs help in deciding which picture-taking magic—also known as imaging studies—is best for those in need?”
So, the researchers set out on a sleigh ride of discovery with two of these clever helpers: ChatGPT (GPT-4) by OpenAI, a jolly fellow, and Glass AI by Glass Health, a rather sharp newcomer. They tested them with 1075 clinical scenarios, much like checking who’s naughty or nice, but for medical cases. These scenarios were from 11 expert panels of the American College of Radiology, who know a thing or two about imaging studies.
Each LLM gave two responses per scenario, and like mixing sugar and spice to get the perfect cookie, they averaged the scores for a final verdict. They even used a special scale from 0 to 3, where partial points were like half-eaten cookies left for partial answers.
And what did they find, you ask? Well, Glass AI was the star on top of the tree, scoring significantly higher than ChatGPT (2.32 vs 2.08, p=0.002). Both did their best work in the Polytrauma, Breast, and Vascular panels—like elves excelling in toy-making, cookie-baking, and reindeer-training. But they stumbled a bit in the Neurologic, Musculoskeletal, and Cardiac panels, much like I sometimes stumble on a snowy rooftop.
Glass AI outshone ChatGPT in nearly all panels, save for OB/GYN, where ChatGPT held its own. The two agreed most on Pediatrics, Neurologic, and Thoracic panels, but had a bit of a snowball fight over Vascular, Breast, and Urologic panels.
In the end, the tale tells us that these LLMs, especially Glass AI with its extra medical-text training, could very well help in choosing the right imaging study, like a compass guiding Santa’s sleigh. It’s a merry thought that these clever creations could one day assist in radiologic decision-making, bringing cheer to patients and doctors alike!
And with that, my elves, let’s get back to our toy-making, for there’s always joy in helping others, be it with a toy or a well-chosen imaging study. 🎄✨